Analysis of univariate linear, robust-linear, and non-linear machine learning algorithms for satellite-derived bathymetry in complex coastal terrain

Abstract

The blue economic, maritime, and exploration activities rely on precise knowledge of coastal bathymetry. The recent trends in Satellite-Derived Bathymetry (SDB) research have focused on various methods of estimation using very high-resolution satellite imagery and in-situ data, but mostly in clear and transparent water. The Indian coastal region mostly has murky water, which is further constrained to employ SDB techniques in areas of river mouth and delta due to the presence of numerous underwater rocks and rich sediment carried by rivers. This study is focused on analyzing SDB in two study areas characterized as turbid, sediment-laden, and complex due presence of numerous underwater rocks. The objective of the research is to analyze and compare the performance of a few univariate Machine Learning (ML) regression algorithms in SDB estimation using ASTER, LANDSAT-8, and Sentinel-2A spectral bands with high resolution in-situ bathymetric data. This study evaluates linear, three robust linear Huber, RANSAC & ThielSen, and a non-linear Gaussian Process Regression ML algorithm, to analyze the efficacy in SDB estimation using a univariate spectral bandwidth of satellite imagery. The applied non-linear regression model has estimated SDB with the accuracy of R2 0.87, RMSE 1.77 m, and MAE 1.27 m for depth of 30 m in site A; and site B, R2 0.91, RMSE 1.51 m & MAE 1.17 m for depth of 22 m. The advantage offered by this approach includes; minimum required parameters, less processing time, and the potential to be an alternative to hydrographic surveys.